A Computer Vision-Based Attention Generator Using DQN

Jordan Chipka, Shuqing Zeng, Thanura Elvitigala, Priyantha Mudalige; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2942-2950

A significant obstacle to achieving autonomous driving (AD) and advanced driver-assistance systems (ADAS) functionality in passenger vehicles is high-fidelity perception at a sufficiently low cost of computation and sensors. An area of research that aims to address this challenge takes inspiration from human foveal vision by using attention-based sensing. This work presents an end-to-end computer vision-based reinforcement learning (RL) technique that intelligently selects a priority region of an image to place greater attention to achieve better perception performance. This method is evaluated on the Berkeley Deep Drive (BDD) dataset. Results demonstrate that a substantial improvement in perception performance can be attained - compared to a baseline method - at a minimal cost in terms of time and processing.